Searching for Robust Augmentations to Improve Out-of-Domain Generalization in Dermoscopic Skin Cancer Classification

This paper searches single augmentations, photometric combinations, and composite policies for a binary malignant-versus-non-malignant dermoscopic classifier. Using ConvNeXt-Large trained on multi-source ISIC Archive data with Derm7pt, it evaluates on predominantly source-disjoint HAM10000 and ISIC 2019-2020 data. The mix policy produced the largest reported OOD improvement: +0.053 ROC-AUC on an expanded held-out-source pool, with a 95% CI of +0.045 to +0.061 and p<0.001. Across four seeds, ROC-AUC improved from 0.761-0.775 to 0.806-0.829. However, augmentation selection reused evaluation sources, and a small clinical sensitivity gain was based on only 22 malignant cases and was not seed-stable.